The free-words problem

The promise was that AI would take work off your plate. For a lot of tasks, it does. What is less discussed is that it quietly moves a different kind of work onto your plate, and that work is harder to see, because it does not look like work.

When you hand a task to a model, you are not finished. You still have to read what came back, judge whether it is right, and decide whether to use it. Researchers call the first half of that trade cognitive offloading: pushing a mental task to an external system. The catch is that offloading the production of an answer does not offload the responsibility for the answer. You are now a reviewer of work you did not do, which is often more demanding than doing it, because you have to reconstruct enough of the reasoning to know whether to trust it.

A 2025 study from researchers at MIT's Media Lab put language to the longer-term version of this. Participants who leaned on a model to write showed weaker memory of their own output and lower measured brain engagement during the task. The authors used the phrase cognitive debt: the gap accrues quietly, like a balance you are not watching, and comes due later as skill you expected to have and do not.

The second pressure is simpler and more physical. The assistant is always available, and always available cuts both ways. There is no longer a natural stopping point, no closed library, no colleague who has gone home. The field that studies this calls it technostress, and the specific variant here is techno-overload: the sense that the tool lets you, and therefore expects you, to do more, faster, without end. The tool does not get tired. You do.

The individual story is only half of it. The more interesting shift happens in the space between people, because when the cost of producing something drops, the cost of sending it to someone else drops with it.

For most of working life, effort acted as a filter. Writing a long request, assembling a detailed brief, or drafting a careful document took enough time that you did it only when it was worth doing. That friction was annoying, and it was also load-bearing: it rationed how much arrived in other people's inboxes. Remove the friction and you do not just make good requests cheaper. You make all requests cheaper, including the ones that were not worth anyone's time.

This is a familiar pattern, just in a new place. Economists know it as the Jevons paradox: when a resource becomes more efficient to use, total consumption of it tends to rise rather than fall, because the efficiency invites more use. Make writing cheaper and you get more writing, not less reading. The person on the receiving end still has to read, interpret, and respond to every piece, and that cost did not move at all.

There is a second layer. More of what crosses between colleagues has now passed through a model on the way out, and often through another model on the way back in, as the recipient asks their own assistant to summarize it. Researchers call this AI-mediated communication: human intent, routed through a model, then interpreted on the far side, sometimes by a model. Each pass is smoothed, expanded, or compressed, and a little of the original signal is spent each time.

Who captures the slack

Here is the part that explains why two people can use the same tool and have opposite experiences of it.

When AI makes a task faster, someone gets the time back. Who gets it depends on who controls the workload. If you set your own capacity, as an independent operator, a contractor, or anyone who decides how much they take on, the time AI saves becomes genuine slack. You finish sooner and the day is shorter, or you take on one more engagement by choice. The gain lands in your pocket.

If you do not control your own workload, the saved time does not become slack. It becomes the new baseline. The task that took a day and now takes an hour does not give you back the day; it raises the expectation to the thing that fits in the hour, and then the next thing. The efficiency is real, but it is captured upstream, by throughput, by expanded scope, by whoever sets the targets. Same tool, same speed-up, and the person feels pressure where the other felt relief.

A note on a term that gets overused. There is real discussion of workslop, meaning AI-generated output that looks like work but carries no real content, shifting the burden of making sense of it onto the receiver. That is a genuine failure mode, but it is a different problem from the one above, and worth keeping separate. High-quality AI-assisted work creates the same overload purely through volume: you can produce excellent, carefully-checked material far faster than anyone can absorb it. The receiver is swamped either way. Quality is not the variable that saves them. Quantity is.

The oldest version of this problem has nothing to do with computers. Long before search engines, people asked around, whether a neighbor, a relative, or whoever seemed to know, and the answer they got was shaped by whoever they happened to ask. We have always outsourced knowing. What changes across eras is not the outsourcing; it is the direction the source leans.

Search engines lean toward the crowd. Ask a search engine about a symptom and the results tilt toward what the most people clicked, linked, and feared, which is why typing a mild complaint into a search box so reliably returns the alarming version. Researchers named this cyberchondria: the tool does not share your anxiety, but its ranking rewards the fearful answer, and you inherit the lean.

Conversational AI leans the other way. Where search pulls toward the crowd's fear, a model trained on human approval pulls toward your own view. Phrase a question so it signals the answer you are hoping for, and the system is disproportionately likely to hand it back to you, confidently. The lean is not toward the crowd; it is toward you.

The point is not that one tool is honest and another is not. Every tool leans somewhere. The card catalog leaned toward whatever the librarian shelved. The crowd leans loud. The model leans agreeable. A capable person does not look for the tool with no lean, because there isn't one. They learn the lean of whatever they are holding and correct for it. That is a literacy, and it is teachable.

In an enterprise, that literacy stops being a personal habit and becomes governance. If a model's default tilt is to agree with the person prompting it, then every place your organization relies on AI output for a consequential decision inherits that tilt, unless someone has designed a check that does not. The specific, documented version of this tilt, where a model agrees, fabricates, and then resists correction while still sounding cooperative, is a failure pattern we have written up in full. See the Sycophantic Compliance Cascade →

RAIC helps enterprise teams build that correction into their AI programs deliberately, mapping where AI output feeds consequential decisions, and designing verification into those points using the OWASP Top 10 for LLMs mapped to NIST AI RMF, so the lean of the tool is managed rather than inherited.